Define a source-faithful editing contract - #295
conorbronsdon wants to merge 3 commits into
Conversation
Separate candidate matches, justified findings, and authorized edits. Preserve meaning and explicit context exceptions while allowing requested transformations without invented claims. Refs #202. Default rollout remains gated on #201 human adjudication and held-out evaluation. Co-authored-by: Codex <noreply@openai.com>
conorbronsdon
left a comment
There was a problem hiding this comment.
Independent GPT-5.6 Sol high review and coordinating Codex review found no remaining editing-contract blocker at b46b1c33888bc8cdbd3604fbc4b98280b627c696. Source-level instructions now align on scope, fidelity, profile applicability, explicit transformations, protected content, and unresolved/no-op behavior.
All 18 local suites and all four CI checks pass. Cross-family diagnostic calls are running under #201; no human adjudication or held-out gate is claimed. Ready as a review draft, not for default rollout until that gate is satisfied.
# Conflicts: # CHANGELOG.md
…pr295 # Conflicts: # CHANGELOG.md
|
Draft refresh completed at �8826be2c03435bc297a2746418663bfa844563c.\n\n- merged current main c979c6489ec0ac81a77236782ba496da81b24cb without rewriting branch history\n- kept #202 under Unreleased and kept the shipped 3.35.0 notes intact\n- reran both canonical sync scripts at 3.35.0\n- local |
Pattern matches currently become edits too easily, and some style instructions encourage changes beyond ordinary cleanup. This draft defines one editing contract across the canonical skill, pattern guidance, examples, and companion handoffs.
It separates candidate matches, justified findings, and authorized edits; preserves facts, attribution, uncertainty, technical meaning, protected content, and source voice; and allows explicitly requested structure/register changes without inventing evidence or experience. Context skips remain applicability decisions. Clean cleanup can return unchanged text, while unresolved findings and explicit transformations are handled separately.
References #202. Based on
7e88455009da2d8f41b314e56ec767763fc6f379; candidateb46b1c33888bc8cdbd3604fbc4b98280b627c696.Validation: all 18 local npm test suites passed; canonical/generated parity, mutation controls, OpenAI plugin validation, router connections, category counts, and self-scan passed. Independent GPT-5.6 Sol high review found no remaining contract blocker after correcting source-level contradictions. Model-assisted scenario reads are diagnostic, not human evaluation.
Draft / release gate: #201 still requires actual cross-family comparisons, blind human adjudication, and held-out release evaluation. No semantic-quality or rollout claim is made from passing mechanical tests. The four-section output presentation is intentionally left for the separate #203 change.